github地址：https://github.com/roboflow/rf-detr（develop分支，f108f85）

1、创建环境

打开Anaconda Prompt创建（python>=3.9）

conda create --name rfdetr python==3.11 
conda activate rfdetr

安装依赖

pip install rfdetr inference==0.50.0 -i https://mirrors.aliyun.com/pypi/simple/

2、安装rf-detr

1）使用命令行

pip install git+https://github.com/roboflow/rf-detr.git

2）下载源码安装

cd E:\Development\python
E:
git clone https://github.com/roboflow/rf-detr.git

（若提示Could not connect to server，打开电脑 设置-网络和Internet-代理-手动设置代理 点击编辑查看端口，比如7897，命令行执行

git config --global http.proxy http://127.0.0.1:7897

git config --global https.proxy http://127.0.0.1:7897

查看是否添加成功：git config --global -l

再重新clone.

（下载完后清理，否则可能影响本地代码推送：git config --global --unset http.proxy
git config --global --unset https.proxy）

）

cd rf-detr
pip install -e .

3、demo测试

1）执行以下语句：

pip uninstall numpy
pip install "numpy<2"
pip install tokenizers==0.22.0

2）权重文件下载

地址：https://storage.googleapis.com/rfdetr/rf-detr-seg-preview.pt

将rf-detr-seg-preview.pt放到根目录

3）查看rfdetr\detr.py 文件 318行 

是否为以下内容，否则修改（如果不修改推理结果没有mask掩码）：

            if isinstance(predictions, tuple):
                if len(predictions) == 3:
                    predictions = {
                    "pred_logits": predictions[1],
                    "pred_boxes": predictions[0],
                    "pred_masks": predictions[2]}
                else:
                    predictions = {
                    "pred_logits": predictions[1],
                    "pred_boxes": predictions[0],
                    }

4）执行demo.py

import os
import supervision as sv
from inference import get_model
from PIL import Image
from io import BytesIO
from rfdetr import RFDETRSegPreview
from rfdetr.util.coco_classes import COCO_CLASSES
import requests

# Initialize RF-DETR Segmentation model
model = RFDETRSegPreview()
# # Optimize for faster inference
model.optimize_for_inference()
def demo(img):
    # Predict with segmentation masks
    detections = model.predict(img, threshold=0.25)

    # Create labels with class names and confidence
    labels = [
        f"{COCO_CLASSES[class_id]} {confidence:.2f}"
        for class_id, confidence
        in zip(detections.class_id, detections.confidence)
    ]

    # Annotate frame with masks and labels
    annotated_img = img.copy()
    annotated_img = sv.MaskAnnotator().annotate(annotated_img, detections)
    annotated_img = sv.BoxAnnotator().annotate(annotated_img, detections)
    annotated_img = sv.LabelAnnotator().annotate(annotated_img, detections, labels)

    return annotated_img


url = "https://media.roboflow.com/dog.jpeg"
image = Image.open(BytesIO(requests.get(url).content))

if __name__=="__main__":
    annotated_img=demo(image)
    sv.plot_image(annotated_img)
    annotated_img.save("test.jpg")

结果如下：

这里结果有重复